rand.init: Random initialization of factor models
In zdebruine/LSMF: Least squares matrix factorization

Initialize factor models "W" and "H" by sampling values from A, simulating comparable random distributions, or specifying uniform distribution bounds

rand.init(
  A,
  k,
  W.method = "sample",
  H.method = "sample",
  n.sd = 1,
  bounds = c(1e-05, 1),
  seed = NULL
)

`A`	input dgCMatrix
`k`	rank
`W.method`	One of c("sample", "random", "bounded")
`H.method`	One of c("sample", "random", "bounded")
`n.sd`	number of standard deviations from the mean over which to sample values in A (excluding zeros). Applies only to method = c("sample", "random")
`bounds`	sample values between bounds if method = "bounded"
`seed`	never ever not remember to set the seed. It's NULL by default

Methods support positive, non-zero initializations only.

random: samples values from a random normal distribution with the same mean as non-zero values in A and the same standard deviation
sample: samples non-zero values in A within n.sd standard deviations of the mean of non-zero values in A
bounded: samples values from a random uniform distribution bounded between values specified in "bounds"

Values in W and H will automatically be scaled to approximate the distribution of A upon multiplication. Possible scenarios:

W.method and H.method = "sample" or "random": W and H are the square roots of their respective sampled values
W.method and H.method = "bounded" or other given distribution, no normalization is applied
W.method = "sample" or "random" and H.method = "bounded" or other given distribution (or the reverse): H is assigned first and is held constant, values in W are raised to the power of log(mean(A)/mean(H))/log(mean(W)) such that mean(W) * mean(H) approximates mean(A)

Basic usage:

For initialization of most models, "random" should suffice and is the ideal method for fast convergence to an accurate and robust local minima
For one-sided bernoulli factorization, initiate the bernoulli matrix bounded between for example bounds = c(0.4, 0.6) and the non-bounded matrix with "random"
For one-sided multinomial factorization, initiate the multinomial matrix with the multinomial distribution or bounds corresponding to the limits of the distribution, and the non-bounded matrix with "random"
For models where "A" is comprised of many different values, "sample" may be a good option which is truer to the original distribution than a simple normal "random" method

If these initialization methods do not satisfy, LSMF also takes matrices as initializations for "W" and "H".

list of matrices "W" and "H"

zdebruine/LSMF documentation built on Jan. 1, 2021, 1:50 p.m.